[PATCH] D135876: [Instcombine] Scalarize vscale splats of vectorized BinOps

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In D135876#3868467 <https://reviews.llvm.org/D135876#3868467>, @MattDevereau wrote:

> Moved VectorCombine tests that are no longer necessary into InstCombine
> Deleted all logic from VectorCombine, and have instead moved BinOp scalarizing logic into InstCombine. I've only seen this pattern emerge in InstCombine for scalable vectors, so this InstCombine is currently guarded by a scalable vector check.

Aha - I was wondering if the motivating case was just a scalable transform that escaped.

I stepped through some of the examples, and I think we can add a more general fold that will work for both scalable and fixed.

Here's a test example:

  define <4 x i8> @splat_binop_splat(<4 x i8> %x, <4 x i8> %y) {
    %xsplat = shufflevector <4 x i8> %x, <4 x i8> poison, <4 x i32> zeroinitializer
    ; call void @use(<4 x i8> %xsplat)
    %b = add <4 x i8> %xsplat, %y
    %bsplat = shufflevector <4 x i8> %b, <4 x i8> poison, <4 x i32> zeroinitializer
    ret <4 x i8> %bsplat
  }

With fixed vectors, this is reduced via InstCombinerImpl::SimplifyDemandedVectorElts(): the final splat means we don't care about anything but the zero elements of x and y, so the splat of x is irrelevant. But InstCombinerImpl::SimplifyDemandedVectorElts() bails on scalable vectors, and I'm not sure what is needed to relax that limitation. There's also an earlier bail out on all shuffles of scalable vectors in InstCombinerImpl::visitShuffleVectorInst().

But this pattern won't fold even with fixed vectors if we uncomment the extra use of xsplat, so that seems like an opportunity to do better by just matching this set of patterns since it's probably fairly common.
So I think we want to do something like this for any vector:
splat (binop splat(X), Y) --> splat (binop X, Y)
splat (binop X, splat(Y)) --> splat (binop X, Y)
...and that should allow existing folds to reduce your examples to the ideal (scalarized binop) IR.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135876/new/

https://reviews.llvm.org/D135876